New: Beta Coming Soon

Turn agent claims into evidence.

Every claim in the PR checked against what actually happened, not just what the agent said it did. So you know before you approve.

No spam. One email when we open the beta.

From agent claim to evidence-backed verdict

The claim
What the agent said it did, pulled from its own words, not paraphrased.
The record
What actually ran when the change was executed again, independently.
The verdict
What the evidence supports, what it refutes, and what it cannot settle.

[ 0.1 ] The problem

Your agent wrote the code, the tests, and the summary you read before clicking approve.

Merged on the agent’s own say-so

Most companies will never run a frontier model. Nearly all of them are already merging code these agents wrote, on the agent’s word that the tests pass; It writes the code. It writes the tests. It edits the configuration that decides which tests run. Then it writes the summary a human reads. Disclosure can’t stay voluntary, But most tools available in the market currently test the agent’s work using the agent’s own instruments. The seat where independent proof should sit is empty. Provenant is built for that seat.

How Provenant works

Verdict

  • "All existing tests pass."SUPPORTED
  • "No new dependencies were added."CONTRADICTED
  • "This refactor changes no behaviour."UNSETTLED

What you get

Approve without reading every line

Four checks that run before you open the PR, so a skim is enough to sign off.

Provenant reads the PR body and extracts every factual assertion the agent made, quoted exactly, never paraphrased. Anything it cannot classify is listed as unsupported rather than quietly dropped. You see the full set of claims before anything is checked.

  • "All existing tests pass."
  • "No new dependencies were added."
  • "Only the auth middleware was changed."

Every claim is held against evidence gathered independently, not against the agent's own account of its work. Test results come from running the change again. File scope and dependency changes come from the diff itself.

Claim
Evidence
"All existing tests pass."
Running the change again

The summary says auth middleware. The diff also touches payments. Provenant lists every change the agent left out of its description, so scope creep surfaces before it becomes yours.

  • 2 tests removed
  • 1 file outside stated scope
  • 1 dependency added

No score, no confidence percentage, no plausible guess. A claim is supported, contradicted, or left unsettled with the reason named. You never have to work out whether to trust the answer.

  • NO_LOCKFILE
  • SECRET_REQUIRED
  • CLAIM_NOT_TESTABLE

Trust what your agents claim.

Compare agent claims with independently observable evidence before you ship on them.